ChipsThe story, in brief

Yandex Open-Sources YaFF: A Zero-Copy Wire Format for Protobuf With Near-Struct Read Speed

10–20% CPU savings. That's what Yandex's open-source YaFF achieves in production ad systems by reimagining how AI infrastructure serializes data.

Paper-cut illustration of an amber microchip with circuit paths extending into a row of data-center cabinets.
The infrastructure powering AI.AI illustration by KeyNews
The KeyNews take

Why it matters

Yandex's YaFF demonstrates that infrastructure efficiency wins at scale come from optimizing data serialization layers—critical for recommendation systems and inference pipelines handling billions of requests. This matters to builders running large-scale AI workloads where CPU overhead directly impacts margins.

The key facts

7 to know
  1. Yandex open-sources YaFF: zero-copy wire format for Protobuf

  2. Flat Layout achieves read speeds within 1.2× of raw C++ struct performance

  3. Four layout options: Fixed, Flat, Sparse, Dynamic

  4. Production deployment in advertising recommendation system

  5. 10–20% CPU savings reported at scale

  6. Maintains .proto as single source of truth

  7. Flat Layout reads within 1.2× of raw C++ struct performance

Go to the source

MarkTechPostmarktechpost.com

Publisher excerpt: In this article we look at YaFF, Yandex's open-source zero-copy wire format for the Protobuf ecosystem. We keep the .proto file as the single source of truth, changing only how data sits in memory. We walk through its four layouts — Fixed, Flat, Sparse, and Dynamic — and the benchmark where Flat…
Read original report
Back to today's editionMore chips news

The wider picture

View all
Paper-cut illustration of an amber microchip with circuit paths extending into a row of data-center cabinets.
AI illustration by KeyNews
Chips01

Google Adds Cycle-Level Kernel Profiling to XProf

A developer-facing tooling improvement that directly enables better TPU utilization and kernel optimization. Practitioners building custom Pallas kernels can now see exactly where cycles are spent, shifting from guesswork to data-driven tuning.

InfoQ AI/ML
Paper-cut illustration of an amber microchip with circuit paths extending into a row of data-center cabinets.
AI illustration by KeyNews
Chips02

Civo unveils first of 40 planned edge data center sites across UK

Edge compute infrastructure is becoming critical for low-latency AI inference and agentic workloads. Civo's distributed network strategy reflects growing demand for regional AI compute capacity outside centralized cloud zones — a structural shift in how AI workloads are deployed.

ITPro
Paper-cut illustration of an amber microchip with circuit paths extending into a row of data-center cabinets.
AI illustration by KeyNews
Chips03

China reviews dependence on Broadcom switches in data centres

China is auditing its reliance on foreign networking hardware for AI data centers as part of a broader push to build domestic alternatives. This reshapes global compute buildout economics and chip supply chains at a moment when AI capacity is the competitive moat.

Financial Times Technology